A quantum chemical dataset of interacting molecular pairs for chemical reaction studies Authors: Seunghun Jang, Gyoung S Na Published: 2025 ABSTRACT AbstractUnderstanding molecular interactions beyond single-molecule properties is critical for studying real-world chemical systems. Quantum chemical calculations of molecule–molecule interactions are computationally demanding, making large, publicly available datasets scarce. Here, we present an efficient framework for generating initial configurations of molecular interaction systems and construct a molecular interaction dataset, containing 49,620 individual molecules and 247,741 molecular pairs spanning chromophore–solvent, solute–solvent, and drug–drug interactions, each associated with experimentally characterized equilibrium structures. Our dataset can be used for theoretical studies and machine learning applications in chemical sciences, particularly for modeling intermolecular interactions and structure-based prediction of experimental properties. In future work, we plan to expand the dataset to include non-equilibrium structures and atomic forces, thereby broadening its applicability to reaction modeling and force field development. Introduction The GDB dataset, which captures the diversity of chemical structures of small organic molecules that meet simple chemical stability and synthetic feasibility rules, has been extensively studied. This series of datasets presents all theoretically possible organic molecular structures within a constrained chemical space, considering their potential usability in materials data science [1–5]. Building on the GDB framework, the QM7 dataset was introduced as a quantum machine dataset, featuring various quantum chemical and mechanical properties such as energy and dipole moment for a subset of molecules (7165 molecules) from GDB-13 [6]. QM7, which integrates the structural diversity of the GDB dataset with computational quantum chemistry characteristics, has been widely used by researchers across multiple disciplines. Expanding on this foundation, the QM9 dataset was generated through quantum chemical calculations of 134,000 small molecules included in GDB-17 [7]. Compared with QM7, the QM9 dataset is characterized by greater molecular complexity owing to the more extensive variety and number of heavy atoms, as well as an increased maximum number of molecular components. Additionally, QM9 encompasses a broader range of molecules and provides a more extensive set of physical and computational properties. Consequently, the QM9 dataset has been widely adopted across diverse fields, including chemistry, material sciences, medicine, dataset, and artificial intelligence [8–12]. Notably, in real-world applications, the physical and chemical properties of most organic molecules are often governed by molecular interactions. Although the properties of a single molecule in an isolated system may be similar to those observed in an interacting molecular system, they are never identical. Therefore, the influence of molecular interactions on chemical properties is significant and cannot be overlooked. A dataset of interacting molecular pairs, involving two or more molecules, is thus essential for data-driven research of chemical behaviour. However, constructing such datasets is challenging, as it requires accounting for numerous initial molecular geometries before optimizing their spatial configurations. Park et al. constructed an experimental dataset that captures the optical property changes owing to interactions between organic solvents and chromophores and reported the features of a graph convolutional neural network (GCN) model that can predict optical properties considering the interactions [13, 14]. However, this GCN approach predicts the interaction outcomes using the characteristics of only two individual molecules, without addressing how the luminescent molecule interacts with the organic solvent. This highlights the need for a molecular-pair dataset to decode molecular interactions. In addition, Truhlar et al. published a dataset of 3037 experimental free energies of solvation or transfer energies for 790 unique solutes in 92 solvents, and Wishart et al. reported a dataset containing medical targeting information for two drug molecules [15, 16]. In this study, we developed an efficient and systematic framework for generating the initial configurations of interacting molecular systems. We conducted large-scale quantum chemical calculations using the abovementioned public datasets of SMILES-based molecular-pair combinations. Based on this framework, we constructed the Interacting Molecular Pairs (IMP) Dataset, a computational quantum chemistry dataset for molecular pairs considering various geometric configurations for chromophore–solvent, solute–solvent, and drug–drug molecule combinations [17]. To overcome the challenges associated with structure generation, attributable to the spatial configuration of two or more interacting molecules, we established the initial structures by randomly selecting 10 possible positions for organic solvent molecules around the chromophore. Through careful structural optimization, we generated optimized molecular-pair structures and extracted quantum chemical properties such as total energy, highest occupied molecular orbital (HOMO) energy, lowest unoccupied molecular orbital (LUMO) energy, HOMO-LUMO gap, and interaction energies. While not included in the primary data attributes, gradient norm information, which can help determine the energy convergence of the optimized structures, is also provided in the CSV file. In contrast, our primary goal was to generate an equilibrium molecular-pair dataset that captures valid probability distributions of optimized structures (10 conformations per pair) that can be directly compared with experimental measurements. While the calculations may be somewhat less accurate, this probabilistic representation is particularly valuable for machine learning applications. For example, state-of-the-art AI models such as MatterGen [18] have demonstrated success by learning complex probability distributions of material and molecular structures and generating new structures based on these distributions. In the same spirit, our intention was to create a dataset with a meaningful probabilistic distribution of equilibrium molecular-pair structures, which can serve as a robust foundation for training large-scale foundation models. Unlike previous molecular-pair datasets [19–25], which mainly contain non-equilibrium structures calculated via one-shot non-self-consistent field (non-SCF) calculations, the IMP dataset provides equilibrium dimer structures obtained through self-consistent field (SCF) geometry optimizations. Each dimer is associated with experimentally characterized molecular pairs, allowing direct correlation between calculated and experimental properties. For each molecular pair, 10 distinct equilibrium configurations are provided, enabling more accurate structure-based prediction of experimental properties. The dataset also includes detailed quantum chemical properties and atomic geometries for all dimers, enhancing its applicability for machine learning studies focused on molecular interactions. These features highlight the novelty and practical relevance of the IMP dataset compared to existing datasets. Additionally, the dataset includes optimized structures and quantum chemical properties for individual molecules, which are required to generate pair molecular structures. Specifically, our dataset provides optimized structures and corresponding quantum chemical properties of 38,924 individual molecules and 194,381 molecule pairs for chromophore–solvent interactions; 6038 individual molecules and 30,157 molecule pairs for solute–solvent interactions; and 4658 individual molecules and 23,203 molecule pairs for drug–drug interactions. Table 1 summarizes the number of optimized structures for each category and their quantum chemical properties. Table 1Number of optimized structures and quantum chemical (QC) properties by data typeIndividual moleculesMolecular pairsDataset type# of optimized structures(xyz file type)# of QC properties(csv file type)# of optimized structures(xyz file type)# of QC properties(csv file type)Chromophore–solvent38,924155,696194,381971,905Solute–solvent603824,15230,157150,785Drug–drug465818,63223,203116,015Total49,620198,480247,7411,238,705 Number of optimized structures and quantum chemical (QC) properties by data type Figure 1 shows examples of molecular-pair structures (top) and their corresponding interaction energies (bottom) from the chromophore–solvent subset of the dataset. The solvents (yellow arrows) are positioned at various locations around the chromophore. Different spatial arrangements between the chromophore and solvent (10 in total) result in varying interaction energies. Specifically, the interaction energies, calculated as the difference in total energy between the molecular pair and its individual molecules, exhibit a range of values. Examples of optimized molecular-pair structures (top) and their corresponding interaction energies (bottom) for the chromophore–solvent subset of the dataset. Shown here are 10 random placements of a solvent molecule around a chromophore, optimized to equilibrium structures, with quantum chemically calculated interaction energies. These examples illustrate how diverse spatial configurations result in different interaction energies (N = 10) The proposed dataset is expected to be of value for structure-based molecular interaction prediction and structure-based prediction of experimental properties, as well as interaction-driven optimized structure modelling, supporting machine learning studies on equilibrium molecular interactions. We believe our dataset represents a significant breakthrough for advancing the study of molecular interactions. Comparison with existing pairwise datasets Several benchmark datasets have been developed to study non-covalent interactions, and they vary significantly in terms of scale, chemical diversity, and available properties. Here, we compare the IMP dataset with representative pairwise datasets, highlighting these aspects. Scale. The number of molecular pairs and structures differs significantly across datasets. For example, the ANI-1x [20] dataset contains ∼5 million non-equilibrium structures, though only a small subset of homodimers is included. DES370K [21] includes 3691 heterodimers and 370,959 non-equilibrium conformations. SPICE [22] expands on DES370K with 345,676 non-equilibrium heterodimer conformations. QDπ [23] combines SPICE heterodimers and ANI-1x homodimers with additional perturbations, while OMol25 [25] covers both homo- and heterodimers derived from biomolecule fragments. Classical benchmark sets, such as S66 [19] and S66×8, are smaller, comprising 66 equilibrium structures and 528 non-equilibrium conformations, respectively. The IMP dataset comprises 247,741 equilibrium structures, encompassing chromophore–solvent, solute–solvent, and drug–drug interactions, with 10 conformers per pair. Diversity. Datasets differ in molecular composition, type of interactions, and conformational sampling. ANI-1x primarily contains small organic molecules (C, H, N, O). DES370K focuses on closed-shell heterodimers, while SPICE adds drug-like molecules and peptides with additional elements. QDπ further diversifies by including homodimers and applying MD simulations and active learning. OMol25 targets biomolecular fragments. S66, S66×8, and GMTKN55 [24] dimer subsets represent canonical non-covalent interactions, such as hydrogen bonding, dispersion, and π–π stacking. IMP dataset uniquely provides experimentally characterized molecular pairs, covering chromophore–solvent, solute–solvent, and drug–drug systems with 10 distinct equilibrium configurations per pair. Availability of properties. The level of quantum chemical detail varies across datasets. ANI-1x provides energies, forces, and multipole moments, but no interaction energies for dimers. DES370K reports interaction energies at the CCSD(T)/CBS level, but no forces or other electronic properties. SPICE and QDπ include energies, interaction energies, forces, multipole moments, and partial charges. OMol25 provides comprehensive properties for biomolecular dimers. S66, S66×8, and GMTKN55 dimer subsets offer high-level benchmark energies. IMP dataset delivers equilibrium geometries, interaction energies, HOMO/LUMO energies, HOMO-LUMO gaps, and saperated atomic geometries for all dimers, supporting direct comparison with experiments and machine learning applications. Overall, IMP dataset combines large scale, diverse molecular coverage, and comprehensive quantum chemical properties for equilibrium molecular pairs, distinguishing it from existing datasets and providing a robust foundation for data-driven studies of molecular interactions. Construction and content Generating molecular-pair data for computational quantum chemistry involves two primary steps. First, structural optimization is performed to obtain a stable structure for each organic solvent or chromophore molecule. Second, a bundle of molecular-pair structures is generated by combining the optimized individual molecules (e.g., chromophore and solvent, solute and solvent, or first and second drug molecules). These molecular pairs then undergo an additional optimization process to refine their structures. Figure 2 illustrates this process, from 3D structure generation to atomic geometry optimization for individual molecules. First, SMILES information (24,810 pairs) for individual molecules was extracted from the Excel file provided in Ref. [13]. Next, the SMILES strings were converted into mol format using the MolFromSmiles function, and hydrogen atom termination was achieved using the AddHydrogens function, both in the RDKit Python package [26]. After converting the 2D molecular structures into 3D structures using the EmbedMolecule function in RDKit, the 3D structures were roughly optimized through energy minimization based on the Merck molecular force field using the MMFFOptimizeMolecule function in RDKit. These 3D structures were saved in the XYZ format. Structural optimization of individual molecules was performed using the ORCA code with the B3LYP functional and the def2-SVP basis set [27–30]. The ORCA input files used for structural optimization are provided in the Supplementary Information (Figure S1). For the chromophore–solvent dataset, the filenames (ID information) of ORCA-optimized structures were aligned with the Tag information reported in Ref. [13], ensuring consistency and ease of use within the dataset. For reference, multiple results for individual molecules with the same molecular structure may exist in the dataset, even if the molecular calculation IDs differ. If the conversion from SMILES to mol failed at Step 3 in Fig. 2 or if the structural optimization process using ORCA (Step 7) was interrupted, these cases were excluded from the dataset. Finally, 49,620 individual molecular calculation datasets were generated (Table 1). Overall process from 3D structure generation to geometry optimization for individual molecules We generated the initial molecular-pair structures using the 49,620 individual molecular calculation datasets, following the workflow described in Fig. 3. The process is described using the chromophore–solvent dataset as an example. First, we selected a sphere with a radius of Rchromophore, capable of encompassing the chromophore molecule. Similarly, we defined a sphere with a radius of Rsolvent for the solvent molecule. Workflow for generating initial molecular-pair structures using individual chromophore and solvent molecule structures. In this example, 10 random placements of the solvent around the chromophore are generated to ensure diverse initial configurations Based on these radius values, we established a sphere with a radius of Rextended, defined as the sum of Rsolvent and Rchromophore, which could sufficiently surround the chromophore molecule and prevent collisions between molecules. Next, we randomly selected 10 coordinates on the surface of this extended sphere. Using these coordinates, we generated 10 initial molecular-pair structures by positioning the centre of the solvent molecule at each of the 10 random surface coordinates. No explicit adjustment of molecular orientations was applied during this step; instead, relative orientations were naturally optimized during geometry optimization. This process resulted in a total of 247,959 molecular-pair structures. To obtain the final optimized solvent-chromophore molecular-pair structures, we performed structural optimization calculations using the ORCA code (Figure S2). During the optimization process, the energy convergence criterion was set to below 1 × 10-5 Eh, and the force convergence criterion was 1 × 10-3 Eh/Å. Excluding the cases (218 molecular pairs) where calculations were interrupted during optimization, we obtained optimized structures for 247,741 pairs. Overall, this molecular-pair dataset included structures with various configurations, reflecting intermolecular interactions, derived from careful optimization calculations. Moreover, only neutral molecules were included in our dataset. Including ions would greatly increase the number of possible molecular-pair combinations and substantially raise the computational cost of geometry optimizations. Therefore, ionic species were excluded from this study. Additionally, to evaluate the practical feasibility of the generated molecular-pair structures, we used a mathematical approach to explore the possible spatial arrangements that multiple molecules can adopt. Figure 4a illustrates potential binding positions, represented by red spheres on the surface of a sphere divided into 8 or 32 segments. The large sphere denotes the central molecule, while the variable s represents the number of available sites on its surface for small molecules to bind. We assumed that s increases with the physical size of the central molecule and calculated the number of possible combinations using the combinatorial formula sCn. Here, n represents the number of small molecules arranged on the surface. The results are shown in 4b. As expected, sCn increases exponentially with even a slight increase in n. a Possible binding sites (red spheres) for small molecules on a central sphere divided into 8 or 32 regions. b Combinatorial growth of configurations based on the number of potential binding sites (s) and number of small molecules arranged on the surface (n), calculated as sCn. c Number of combinations increases sharply, even for small n (e.g., 1 or 2) values Figure 4c shows the sCn values when n is 1 or 2. When s = 32 and n = 2, there exist 496 possible combinations. In other words, even when considering only combinations in which two small molecules are positioned around a moderately large central molecule, a large number of cases must be considered. Therefore, we conducted quantum chemical calculations only for one-to-one molecular combinations, in which a central molecule binds to a single small molecule, consistent with prior work research [31–33]. We have organized and published the Python code for the process introduced in this section as a framework for generating and optimizing molecular-pair structures (https://github.com/SeunghunJang/MPSGO). This framework provides all codes for the processes adopted in this study, including the generation of individual molecule structures, optimization of individual molecules, generation of molecular-pair structures using the optimized individual molecules, optimization of the generated molecular pairs using a quantum chemical calculation code, and extraction of quantum chemical properties. Users can apply this framework to generate new molecular-pair datasets using their own molecule structure inputs. The IMP dataset includes relaxed atomic geometries (XYZ type) for individual molecules and molecular pairs, as well as quantum chemical properties extracted from the structural optimization results [17]. The dataset for individual molecules contains the optimization results for both the chromophore and solvent (or solute and solvent, or drug–drug pairs). To enhance user convenience, we have synchronized the ID information in our dataset with the Tag information in Ref. [13], enabling seamless integration with existing experimental data. The dataset includes quantum chemical properties (in CSV format), including total energy, HOMO-LUMO gap, HOMO energy, and LUMO energy, extracted from structural optimization calculations for individual molecules. Additionally, our dataset includes classification numbers, which clearly distinguish between solvents and chromophores (or solutes and solvents, or drug–drug pairs), and gradient norm values, which can be used to assess the reliability of the structure optimization results. Here, Eh represents the Hartree energy unit, and a0 denotes the Bohr radius. The quantum chemical property data format for molecular pairs in our dataset is nearly identical to that of individual molecules. However, for molecular pairs, the extracted quantum chemical properties also include the interaction energy arising from molecule-molecule interactions. The interaction energy is defined as Emolecular-pair − (E1st-molecule + E2nd-molecule), where Emolecular-pair is the total energy of the physically bound molecular pair, and E1st-molecule and E2nd-molecule denote the total energies of isolated first and second molecules, respectively. As each molecule-molecule combination has 10 different configurational geometries, the data of molecular pairs include more extensive structural optimization results than the atomic structures available for individual molecules. Tables 2 and 3 present the names, units, data types, and brief descriptions of the quantum chemical properties for individual molecules and molecule pairs, respectively. Examples of a relaxed chromophore structure (mol_12742_1.xyz) and molecular-pair structure (12742_5.xyz) are presented in Tables 4 and 5, respectively. Table 2Description of quantum chemical properties dataset for individual moleculesNoColumn nameUnitData typeDescription1ID–IntegerNumber of data points2CLASSIFICATION NUMBER–IntegerChromophore is 1 and solvent is 23XYZ NAME–StringXYZ file name for relaxed structure4TOTAL ENERGYEhFloatTotal energy5HOMO-LUMO GAPeVFloatHOMO-LUMO gap6HOMOeVFloatHOMO energy7LUMOeVFloatLUMO energy8GRADIENT NORMEh/a0FloatGradient norm Description of quantum chemical properties dataset for individual molecules Table 3Description of quantum chemical properties dataset for molecular pairsNoColumn nameUnitData typeDescription1ID–IntegerNumber of data points2CONFIGURATION NUMBER–IntegerStructural number in specific combination3XYZ NAME–StringXYZ file name for relaxed structure4TOTAL ENERGYEhFloatTotal energy5HOMO-LUMO GAPeVFloatHOMO-LUMO gap6HOMOeVFloatHOMO energy7LUMOeVFloatLUMO energy8GRADIENT NORMEh/a0FloatGradient norm of relaxed structure9INTERACTION ENERGYEhFloatInteraction energy (Eh)10INTERACTION ENERGYeVFloatInteraction energy (eV) Description of quantum chemical properties dataset for molecular pairs Table 4Quantum chemical properties of relaxed chromophore structure (mol_12742_1.xyz)ID12742CONFIGURATION NUMBER1XYZ NAMEmol_12742_1.xyzTOTAL ENERGY (Eh)- 53.0201HOMO-LUMO GAP (eV)2.5595HOMO (eV)- 11.0638LUMO (eV)- 8.5043GRADIENT NORM (Eh/a0)0.000363 Quantum chemical properties of relaxed chromophore structure (mol_12742_1.xyz) Table 5Quantum chemical properties of relaxed molecular-pair structure (12742_5.xyz)ID12742CONFIGURATION NUMBER5XYZ NAME12742_5.xyzTOTAL ENERGY (Eh)- 72.0708HOMO-LUMO GAP (eV)2.190HOMO (eV)- 10.706LUMO (eV)- 8.516GRADIENT NORM (Eh/a0)0.000565INTERACTION ENERGY (Eh)- 0.000279INTERACTION ENERGY (eV)- 0.007595 Quantum chemical properties of relaxed molecular-pair structure (12742_5.xyz) Discussion Among the quantum chemical properties provided, total energy is a fundamental characteristic directly obtained from the calculation results. Given that the total energy increases in the negative direction as the number of atoms in a molecule increases, it can indicate molecular size. Figure 5 shows the total energy distribution for individual molecules and their molecular pairs. In the chromophore–solvent system, the total energy distribution of chromophore molecules ranges from - 10 to - 200 Eh, with a peak near - 60 Eh. In contrast, solvent molecules exhibit a more concentrated total energy distribution, ranging from 0 to - 40 Eh. These results indicate that chromophore molecules are larger and contain more atoms than solvent molecules. Total energy distributions for individual molecules and molecular pairs. Individual molecules: Chromophore (N = 19,462), Solvent (N = 19,462), Solute (N = 3019), Solvent (N = 3019), Drug1 (N = 2329), Drug2 (N = 2329). Molecular pairs: Chromophore–solvent (N = 194,381), Solute–solvent (N = 30,157), Drug–drug (N = 23,203). Colors: Sky blue (chromophore–solvent), light green (solute–solvent), orange (drug–drug) Additionally, the total energy values for molecule pairs shift slightly further in the negative direction compared with those of chromophore molecules. As a molecular pair consists of a chromophore and solvent, it can be considered a larger molecular entity than a chromophore alone. This trend is clearly reflected in the total energy distribution. Our dataset consistently aligns with the total energy values, a fundamental property in quantum chemistry. As shown in Fig. 5, a similar trend is observed in the solute–solvent and drug–drug datasets, where the total energy distribution of molecule pairs shifts further in the negative direction compared with that of individual molecules. Our dataset also provides interaction energy values, calculated by combining the quantum chemical calculation results for individual molecules and molecular pairs. Figure 6 presents the interaction energy distributions for 194,381 chromophore–solvent pairs. The upper part of the figure shows the interaction energy distribution across the entire frequency range, while the lower part provides an enlarged view of the distribution for frequencies below 900. Interaction energy values are mostly concentrated between 0 and - 0.05 Eh. Although positive interaction energies and values below - 0.05 Eh are extremely rare, our dataset still includes these cases. By leveraging the molecular-pair structures and interaction energy values in our dataset, we can develop an artificial intelligence model using a graph neural network to predict the interaction energy between two different molecular structures at arbitrary spatial positions. As shown in the Supplementary information (Figures S3 and S4), the interaction energy distributions for solute–solvent and drug–drug pairs exhibit trends similar to those of the chromophore–solvent pairs. Distribution of interaction energies for chromophore–solvent pairs (N = 194,381) This study is aimed at providing optimized molecular structures obtained through quantum chemical calculations, along with their corresponding quantum chemical properties, to support data-driven research on molecular interactions. To this end, the reliability of these optimized structures must be verified. As previously mentioned, the gradient norm provides a measure of how far a molecular structure is from reaching its optimized state. To illustrate this more intuitively, molecular structures are arranged by their gradient norm values in (Figure S5). As shown in the figure, the gradient norm decreases as the number of geometry optimization cycles increases. When the gradient norm decreases from 0.41 to 0.03 Eh/a0, structural changes are minimal; however, a noticeable structural shift occurs when the value drops below 0.0028 Eh/a0. A gradient norm of 0.00025 Eh/a0 indicates a fully optimized structure. Figure 7 illustrates the distribution of gradient norms for the optimized individual molecules and molecular pairs provided in our dataset. The gradient norms for the chromophore and solvent molecules are mostly below 0.001 Eh/a0, indicating that the provided structures are well optimized. In the case of molecular pairs, only a tiny fraction of the data (N = 172, less than 0.09%) exhibits values above 0.004 Eh/a0. Thus, most of the dataset (more than 99.9%) consists of reliably optimized structures. As noted in the discussion on generating molecular-pair configurations above, our optimized structures may not necessarily represent global minima. Instead, they encompass a diverse set of local minimum structures, offering possible candidate configurations for various solvent molecules surrounding the chromophore. Gradient norm distributions for optimized individual molecules and molecular pairs, including chromophore molecules (N = 19,462), solvent molecules (N = 19,462), chromophore–solvent molecular pairs (N = 194,209), and the subset of chromophore–solvent molecular pairs with gradient norms exceeding 0.004 Eh/a0 (N = 172) To further examine the influence of these configurations, the density-based distributions of the HOMO–LUMO gap for individual (chromophore and solvent) molecules and for molecular pairs are shown in Figure S6. The figure demonstrates that the HOMO–LUMO gaps of the molecular pairs are primarily determined by the chromophore components. Notably, the HOMO–LUMO gaps in molecular pairs are generally smaller than those in individual chromophores, reflecting the effect of the surrounding solvent. This trend is consistently observed throughout the dataset, providing evidence that our dataset captures meaningful interactions between chromophores and solvent molecules. Conclusion In this work, we presented the IMP Dataset, a comprehensive quantum chemical dataset designed to capture the effects of molecular interactions on chemical properties. While existing datasets such as QM7 and QM9 have significantly advanced the field of machine learning in quantum chemistry by providing molecular-level quantum chemical properties, they primarily focus on isolated molecules. However, in real-world chemical and biological systems, molecular interactions play a crucial role in determining physical and chemical behavior. This underscores the necessity for a dataset that explicitly considers pairwise molecular interactions. To address this gap, we developed an efficient and systematic framework for generating initial configurations of interacting molecular systems, followed by large-scale quantum chemical calculations. The IMP dataset comprises optimized geometries and quantum chemical properties of both individual molecules and molecular pairs across various categories, including chromophore–solvent, solute–solvent, and drug–drug interactions. By considering diverse spatial configurations and molecular types, this dataset supports more detailed studies of interaction-driven quantum chemical phenomena. The IMP dataset is expected to serve as a valuable resource for applications, including structure-based prediction of molecular interactions and experimental properties, interaction-driven structural modeling, and machine learning studies on equilibrium molecular interactions. Limitations and future perspectives While the IMP dataset provides optimized structures and energies for 10 distinct equilibrium configurations of each molecular pair (heterodimer), it has several limitations. Specifically, the dataset is currently restricted to static geometries of pairwise interactions and does not incorporate dynamic sampling, thermal effects, or higher-order interactions. Nevertheless, the IMP dataset should be viewed as a valuable starting point for studies of molecular interactions and related experimental properties. The comprehensive information on interacting molecular pairs allows it to serve as a base dataset from which new datasets can be generated. For example, it could be extended to construct higher-order molecular systems involving three or more molecules, or to generate large-scale non-equilibrium datasets for chromophores, drug molecules, and other molecular systems through additional geometric perturbations, molecular dynamics simulations, and active learning techniques. Future work is expected to build upon this foundation by incorporating ternary or higher-order systems and molecular dynamics simulations, thereby providing a more comprehensive representation of molecular environments and interactions. Additional file Supplementary file 1. Author contributions S.J. designed and initiated the project, generated and analyzed the dataset, and wrote the original manuscript. S.J. and G.S.N. jointly supervised the project and revised the manuscript. Funding This research was funded by the Ministry of Trade, Industry and Energy (No. TS251-10R) and Korea Research Institute of Chemical Technology (No. KK2551-10). Data availability The Python framework for generating the IMP dataset and our dataset is available at https://github.com/SeunghunJang/MPSGO and https://doi.org/10.6084/m9.figshare. 29105159. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary Information The online version contains supplementary material available at 10.1186/s13321-025-01124-y. References Associated Data Supplementary Materials Supplementary file 1. Data Availability Statement The Python framework for generating the IMP dataset and our dataset is available at https://github.com/SeunghunJang/MPSGO and https://doi.org/10.6084/m9.figshare. 29105159.